Does MPI_File_write_at() initialize the file with zeros?

Question

Consider the following simple program which writes the rank of all processes whos rank is bigger than zero into a file:

#include <mpi.h>

int main() {
    MPI_Init(NULL, NULL);

    int world_rank, world_size;
    MPI_Comm_rank(MPI_COMM_WORLD, &world_rank);
    MPI_Comm_size(MPI_COMM_WORLD, &world_size);

    MPI_Offset offset;
    MPI_Status status;
    MPI_File fh;

    MPI_File_open(MPI_COMM_WORLD, "myfile", MPI_MODE_CREATE | MPI_MODE_WRONLY,
            MPI_INFO_NULL, &fh);
    
    offset = world_rank * sizeof(int);

    if (world_rank > 0) {
        MPI_File_write_at(fh, offset, &world_rank, 1, MPI_INT, &status);
    }

    MPI_File_close(&fh);

    MPI_Finalize();

    return 0;
}

we compiled and run it on 4 processes

mpic++ main.cpp
mpirun --oversubscribe -n 4 a.out

We check the written file with hexdump -C myfile

00000000  00 00 00 00 01 00 00 00  02 00 00 00 03 00 00 00  |................|
00000010

Now, I never made a write call to the first integer aka the first 4 bytes but they are zero.

Can I be sure that those are always zero?

dreamcrash · Accepted Answer · 2020-12-14T19:55:43.930

MPI_File_write_at will write the data into the file as binary data. Then when you run the command hexdump -C myfile that command will display the data, accordingly, with the first 4 bytes being the offset. Those, first bytes are not part of the binary data per si, but rather added by hexdump -C myfile for readability purposes.

The hexadecimal 00000010 represents 10000 in binary and 16 in decimal. If you look at your first line, ignoring the first 4 bytes,:

  4 bytes   4 bytes      4 bytes      4 bytes       4 bytes 
00000000 | 00 00 00 00  01 00 00 00  02 00 00 00   03 00 00 00
00000010 |

You have 16 (4x4) bytes, hence why the next row starts with 00000010.

Can I be sure that those are always zero?

As far as the standard is concerned, I have not found there, explicitly stating that if one skips the beginning of the file offset > 0, that the MPI implementation will fill that gap with zeros. For instance, with the MPI version (Open MPI 1.8.8) that I am using, if I modify your code to:

 if (world_rank == 3) {
    MPI_File_write_at(fh, offset, &world_rank, 1, MPI_INT, &status);
}

I get the following output from hexdump -C myfile:

00000000  00 00 00 00 00 00 00 00  00 00 00 00 03 00 00 00  |................|
00000010

So with the MPI version that I am using, and apparently with yours, it initializes with zeros.

Nonetheless, unless a reliable source can be found (which I did not managed to find one) that explicitly states that in your case the first 4 bytes will always be zero, I would advice to not make any assumptions on that regard. Notwithstanding, one should not care about the file content part that was not written by processes, anyway.

EDIT: A clarification from "the Open MPI mailing list":

In general, the contents of a file written by the MPI IO interface are going to be implementation-specific.

Does MPI_File_write_at() initialize the file with zeros?

1 Answers1